Predicting the Unpredictable: How Dynamic COVID-19 Policies and Restrictions Challenge Model Forecasts (preprint)

Between June and August 2020, an agent-based model was used to project rates of COVID-19 infection incidence and cases diagnosed as positive from 15 September to 31 October 2020 for 72 geographic settings. Five scenarios were modelled: a baseline scenario where no future changes were made to existing restrictions, and four scenarios representing small or moderate changes in restrictions at two intervals. Post hoc, upper and lower bounds for number of diagnosed Covid-19 cases were compared with actual data collected during the prediction window. A regression analysis with 17 covariates was performed to determine correlates of accurate projections. It was found that the actual data fell within the lower and upper bounds in 27 settings and out of bounds in 45 settings. The only statistically significant predictor of actual data within the predicted bounds was correct assumptions about future policy changes (OR = 15.04; 95%CI 2.20-208.70; p = 0.016). Frequent changes in restrictions implemented by governments, which the modelling team was not always able to predict, in part explains why the majority of model projections were inaccurate compared with actual outcomes and supports revision of projections when policies are changed as well as the importance of modelling teams collaborating with policy experts.

Predicting the unpredictable: how dynamic COVID-19 policies and restrictions challenge model forecasts (preprint)

IntroductionTo retrospectively assess the accuracy of a mathematical modelling study that projected the rate of COVID-19 diagnoses for 72 locations worldwide in 2021, and to identify predictors of model accuracy. MethodsBetween June and August 2020, an agent-based model was used to project rates of COVID-19 infection incidence and cases diagnosed as positive from 15 September to 31 October 2020 for 72 geographic settings. Five scenarios were modelled: a baseline scenario where no future changes were made to existing restrictions, and four scenarios representing small or moderate changes in restrictions at two intervals. Post hoc, upper and lower bounds for number of diagnosed Covid-19 cases were compared with actual data collected during the prediction window. A regression analysis with 17 covariates was performed to determine correlates of accurate projections. ResultsThe actual data fell within the lower and upper bounds in 27 settings and out of bounds in 45 settings. The only statistically significant predictor of actual data within the predicted bounds was correct assumptions about future policy changes (OR = 15.04; 95%CI 2.20-208.70; p=0.016). ConclusionsFor this study, the accuracy of COVID-19 model projections was dependent on whether assumptions about future policies are correct. Frequent changes in restrictions implemented by governments, which the modelling team was not always able to predict, in part explains why the majority of model projections were inaccurate compared with actual outcomes and supports revision of projections when policies are changed as well as the importance of policy experts collaborating on modelling projects.